Computer Science and Artificial Intelligence Laboratory

Technical Report

MIT-CSAIL-TR-2009-039 August 26, 2009

AvatarSAT: AN AUTO-TUNING BOOLEAN

SAT SOLVER
RISHABH SINGH, JOSEPH P. NEAR, VIJAY GANESH,
and MARTIN RINARD

m a ss a c h u se t t s i n st i t u t e o f t e c h n o l o g y, c a m b ri d g e , m a 02139 u s a — w w w. c s a il . mi t . e d u
 AvatarSAT: An Auto-tuning Boolean SAT Solver

Rishabh Singh, Joseph P. Near, Vijay Ganesh, and Martin Rinard

Massachusetts Institute of Technology

{rishabh, jnear, vganesh, rinard}@csail.mit.edu

Abstract. We present AVATAR SAT, a SAT solver that uses machine-learning

classifiers to automatically tune the heuristics of an off-the-shelf SAT solver
on a per-instance basis. The classifiers use features of both the input and con-
flict clauses to select parameter settings for the solver’s tunable heuristics. On
a randomly selected set of SAT problems chosen from the 2007 and 2008 SAT
competitions, AVATAR SAT is, on average, over two times faster than M INI SAT
based on the geometric mean speedup measure and 50% faster based on the arith-
metic mean speedup measure. Moreover, AVATAR SAT is hundreds to thousands
of times faster than M INI SAT on many hard SAT instances and is never more
than twenty times slower than M INI SAT on any SAT instance.

1 Introduction

We present AVATAR SAT, an auto-tuning SAT solver that uses machine-learning meth-
ods to automatically tune an off-the-shelf SAT solver on a per-instance basis. AVATAR -
SAT is designed to classify each SAT instance based on its syntactic features so as to
find the corresponding optimal parameter setting for the heuristics in the underlying
SAT solver (for example, dynamic variable selection [11] and restarts [5]).
AVATAR SAT is built on top of M INI SAT version 2.0, a state of the art SAT solver [5].
Based on its performance in recent SAT competitions [1], M INI SAT can reasonably
claim to be one of the fastest open-source SAT solvers.
We compared AVATAR SAT and M INI SAT on a randomly selected set of SAT prob-
lems chosen from the 2007 and 2008 SAT competitions [1]. AVATAR SAT is, on the
average, more than two times faster per problem than M INI SAT based on the geomet-
ric mean speedup measure and 50% faster per problem based on the arithmetic mean
speedup measure. 1
On many hard SAT instances AVATAR SAT is hundreds to thousands of times faster
than M INI SAT, and is never more than twenty times slower than M INI SAT on any SAT
instance.

Availibility: AVATAR SAT’s source code, experimental data, and results are available at
http://people.csail.mit.edu/vganesh/avatarsat.html.
1
The geometric mean speedup is the geometric mean of the speedups for the individual tests.
The arithmetic mean speedup is the ratio of the total time taken by M INI SAT divided by the
total time taken by AVATAR SAT over all tests.
 Solver Instances Solved Time-outs Total Time Geometric Arithmetic
(in seconds) Speedup Speedup
AVATAR SAT 75 53 22 242,090.48 1.51X 2.23X
M INI SAT 2.0 75 52 23 366,353.35 - -

Fig. 1. Number of problems solved and time taken by AVATAR SAT and M INI SAT.

Contributions:

• AVATAR SAT: AVATAR SAT is a modified version of M INI SAT 2.0 that uses ma-
chine learning classifiers to choose parameter settings for the tunable heuristics that
control several aspects of M INI SAT’s search algorithm.
• Course Correction During Search: A key novelty in AVATAR SAT is the course
correction step. Modern SAT solvers accumulate conflict clauses and drop input
clauses during their search, which can change the structure of the problem consid-
erably. The optimal parameter settings for this new problem may be significantly
different from those of the original input problem.
AVATAR SAT therefore first selects a set of parameters for M INI SAT to use dur-
ing the initial part of the search. When the number of new clauses accumulated
during this part of the search crosses a threshold, the course correction step exam-
ines the new clauses to select a new set of parameters for M INI SAT to use during
the remaining part of the search. In this way AVATAR SAT dynamically adapts the
parameter settings to the potentially changing characteristics of the SAT problem.
• Use of Support-Vector Machines: AVATAR SAT uses multiclass support-vector
machines (SVM) [4], a supervised machine-learning technique, to learn a function
from the features of SAT instances to discretized parameter settings.
• Experimental Results: We present experimental results comparing the perfor-
mance of AVATAR SAT and M INI SAT on 75 randomly selected SAT instances
(both crafted and industrial) from the SAT 2007 and 2008 competitions [1].

2 The Architecture of AVATAR SAT

AVATAR SAT is a modified version of M INI SAT that relies on two machine-learning
classifiers to automatically tune its heuristic search. The architecture of AVATAR SAT is
summarized in Figure 2. The preprocessing classifier is invoked on the input SAT in-
stance to generate initial parameter settings for the solver’s heuristics, while the course
correction classifier is invoked on both the original set of input clauses and the conflict
clauses generated during the search in an attempt to adapt the heuristic search to the
changing problem instance. Both classifiers are built using the well-known LIBSVM
library [3]. The solver itself works as follows:

• Initial Classification: AVATAR SAT computes features of the input SAT instance.
The preprocessing classifier uses these features to select initial parameters for M INI -
SAT’s heuristics.
 AVATAR SAT Number AVATAR SAT M INI SAT 2.0
Speedup Over of Tests Avg. Time Avg. Time
M INI SAT 2.0 (Seconds) (Seconds)
1000–50000x 2 3.41 43,200.00
100–1000x 3 57.50 28,850.61
10–100x 3 109.15 4,588.56
1.5–10x 12 1,106.8 3,707.90
1/10–1/1.5x 9 11,636.95 2,941.76
1/20–1/10x 1 4,332.67 260.18

Fig. 2. AVATAR SAT Architecture Fig. 3. Number of SAT instances in which AVATAR -
SAT exhibits speedups over M INI SAT; AVATAR SAT
never slows down below 1/20x of M INI SAT. The
timeout was 12 hours.

• Heuristic Search: AVATAR SAT invokes the modified M INI SAT with the selected
parameter settings.
• Course Correction: When the conflict clauses accumulated during the solver’s
search reach a pre-determined fraction of the input clauses (80%), AVATAR SAT
invokes the course correction classifier on the input and conflict clauses to select
new search parameters for the remainder of the search.

Machine Learning Technique Used: AVATAR SAT uses the multi-class support vec-
tor machine (SVM) [4] supervised machine-learning technique. Supervised machine-
learning techniques attempt to learn an unknown function based on a training set con-
sisting of input vectors and the corresponding outputs. The SVM technique trains a
classifier that maps features of SAT instances to one of a finite set of classes, where
each class corresponds to a different parameter configuration. Since the generation of
training data requires discretization of parameter values, a multi-class SVM is a natural
choice for classification.
The SAT instance features used in training are designed to characterize the cor-
responding SAT problem instance as closely as possible. They are also designed to
be of reasonable length and to be efficient to compute. Our feature vector consists of
58 different log-normalized features of the input SAT instance (e.g., clause/var ratio,
var/clause ratio, number of variables, clauses etc. In addition to novel features such
as clause-weighted positive and negative literal occurrences for distinguishing shuffled
SAT instances, we also use some of the features in [12]).

MiniSAT Heuristics Chosen for Automatic Tuning: The SAT solver parameters that
are dynamically tuned in AVATAR SAT are the variable decay parameter of the VSIDS
heuristic [11], and restart increments [5]. Of the ten tunable parameters in M INI SAT,
we selected these two because they have been empirically observed to have the most
significant influence on solver performance [2].
 AvatarSAT vs. MiniSAT 2.0 Problems Solved Over Time
100000
40000 MiniSAT 2.0
AvatarSAT
10000 35000
30000
AvatarSAT

1000 25000

Time
20000
100
15000
10000
10
5000
1 0
1 10 100 1000 10000 100000 0 10 20 30 40 50 60
MiniSAT 2.0 Number of Problems Solved

Fig. 4. Log-scale scatterplot of running times Fig. 5. Number of problems solved by

(in seconds) of AVATAR SAT vs. M INI SAT. AVATAR SAT and M INI SAT over time. In
Each point represents at least one test case; our experiments, AVATAR SAT solves more
the timeout was set at 12 hours. problems than M INI SAT for any given
amount of time.

3 Results
We discuss the experimental setup, guidelines used to conduct the experiments, training
setup of the machine-learning classifiers, and results of the comparison of AVATAR SAT
with M INI SAT.
Experimental Setup: All experiments for comparing AVATAR SAT with M INI SAT,
and training of the machine-learning classifiers were conducted on a cluster of 75 Linux
machines each with 1.8GHz Intel Xeon processor, 4MB Cache, and 2 GB RAM. To
train the preprocessing classifier, we randomly split the SAT 2007 and 2008 competi-
tion benchmarks [1] into disjoint training and test sets; the training set contained 177
instances, while the test set contained 75. To train the course correction classifier, we
dumped 391 internal conflict clause instances generated during the execution of M INI -
SAT on the instances used to train the preprocessing classifier. The test set was not
involved in training either classifier; no distinctions between industrial and crafted or
satisfiable and unsatisfiable examples were made.
Experimental Guidelines Used: We followed the guidelines presented by Zarpas [15]
for SAT solver benchmarking experiments. In particular, the timeouts for testing the
solvers were set relatively high at 12 hours (43,200 seconds). Low timeouts are used in
SAT competitions for logistical reasons. However, in real applications solvers may be
given much higher timeouts. Furthermore, higher timeouts help differentiate one solver
from another on harder examples. In our speedup computation, we excluded those test
cases for which both M INI SAT and AVATAR SAT timed out.
Training The Classifiers: The first step in the training process is collection of raw-
data: We ran different configurations of M INI SAT on the SAT instances in our training
set. Each configuration is a pair of parameter settings2 for the variable decay and restart
2
VAR-DECAY ∈ {0.50, 0.75, 0.85, 0.91, 0.93, 0.95, 0.97, 0.99, 0.999}
RESTART-INC ∈ {1, 1.25, 1.5}
increment parameters of M INI SAT. For each instance in the training set, M INI SAT
was invoked with 27 different configurations (9 different values for variable decay, and
3 for restart-increment) and a timeout of 900 seconds per configuration. For each train-
ing instance, and for each M INI SAT configuration, the corresponding runtimes were
recorded. The raw data for each training instance was compiled into a training exam-
ple comprising the features of the instance and the parameter configuration for which
M INI SAT finished fastest. The preprocessing classifier was trained using the training
examples compiled from the 177 training instances discussed above, while the course
correction classifier was trained using the examples compiled from the 391 conflict
clause training instances. Both classifiers were trained using the Radial Basis Function
(RBF) Kernel [3].

3.1 Comparison of AVATAR SAT with M INI SAT

Figure 4 is a scatterplot (in log-scale) of the times taken by AVATAR SAT (Y-axis) and
M INI SAT (X-axis) over all the 75 SAT instances in the test suite. There is one point
for each solved SAT instance. For each point, the ratio of the X-axis to the Y-axis is the
speedup of AVATAR SAT over M INI SAT for the corresponding SAT instance. AVATAR -
SAT is faster for points below the diagonal; M INI SAT is faster for points above the
diagonal.
These data show that AVATAR SAT is faster than M INI SAT for 34 of the 75 in-
stances, with M INI SAT faster on 24 others. For the remaining examples both AVATAR -
SAT and M INI SAT time out without solving the instance. AVATAR SAT is hundreds to
thousands of times faster than M INI SAT 2.0 on 10 hard SAT instances out of a total
of 75 instances (note that some of the corresponding points in the lower right corner
overlap). On at least 5 of these examples M INI SAT times out at 43,200 seconds, while
AVATAR SAT finishes relatively quickly in a few seconds with the correct answer. A
summary of the distribution of differences in solving time is given in Figure 3.
We attribute the fact that M INI SAT is sometimes faster than AVATAR SAT to clas-
sification errors on the part of the machine learning classifiers. Note, however, that
these classification errors never cause AVATAR SAT to execute more than twenty times
slower than M INI SAT, with the vast majority of the misclassifications resulting in
much smaller performance differences between the two solvers. The end result is that
AVATAR SAT is on average more than two times faster per problem than M INI SAT
based on the geometric mean speedup measure, and 50% faster based on the arithmetic
speedup measure.
Figure 5 shows the number of examples solved (X-axis) as a function of the cumu-
lative solution time (Y-axis). At every point in time, AVATAR SAT has solved at least as
many problems as M INI SAT. Note that, as shown in Figure 1, AVATAR SAT solves 53
SAT instances out of a total of 75, compared to 52 by M INI SAT. Moreover, AVATAR -
SAT solves these instances much faster.
We also performed experiments with only the initial classification step (i.e., the
course correction step was disabled so that the solver used the initially selected param-
eters for the entire search). The course correction step was found to be important for
obtaining good performance on some of the harder instances. It usually makes little to
no difference for examples that solve quickly. This reflects the substantial difference in
problem characteristics that can arise between the original problem and the set of de-
rived conflict clauses that the solver accumulates as it attempts to solve a hard instance.
Note that even though the classifiers were trained with data from executions with
small timeouts, they generalize well to tests with much larger timeouts.

4 Related Work

There has been recent work in combining machine-learning preprocessors with SAT
solvers for the purposes of predicting parameter values and for portfolio selection. In
portfolio selection, a classifier predicts the best solver from a fixed set for a particular
SAT instance [12, 14, 6, 13]. Previous work in predicting parameter values has used lin-
ear regression classifiers to learn a function from features and parameter configurations
to solver running time [8]. Other work has focused on per-distribution tuning of param-
eters [7, 9], in which a fixed set of parameters is chosen for a given distribution based
on average running time.
All of these techniques use some kind of a classifier as a pre-processor. By contrast,
AVATAR SAT calls a machine-learning classifier as a preprocessor, and another classi-
fier internally to correct the course of the solver’s search. Additionally, the strategy of
learning a function from features and parameter configuration to runtime does not scale
as well as our approach of learning a function from features to parameter configura-
tion. The former strategy requires invoking the classifier once per possible parameter
configuration—a number of invocations possibly exponential in the number of param-
eters. Our strategy, by contrast, requires invoking the classifier only once, regardless of
the number of parameters.
The use of machine learning techniques to dynamically adjust the search strategy
of a SAT solver has been examined [10], but existing approaches are neither as general
nor as effective as ours. These approaches have not been tested on the kind of varied
problem sets represented by those collected for the SAT competitions, and even so
achieve only modest performance improvements. They also rely on branching rules
specific to the contemporary SAT solvers of several years ago—rules that have since
become obsolete. Our technique, in contrast, allows the tuning of any parameterizable
feature of a SAT solver, and so will remain relevant even as solvers improve.

References

1. SAT competition website. http://www.satcompetition.org/.

2. G. Audemard and L. Simon. Experimenting with small changes in conflict-driven clause
learning algorithms. In Proceedings of the 14th international conference on Principles
and Practice of Constraint Programming (CP), pages 630–634, Berlin, Heidelberg, 2008.
Springer-Verlag.
3. C. Chang and C. Lin. LIBSVM: a library for support vector machines, 2001. Software
available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm.
4. C. Cortes and V. Vapnik. Support-Vector Networks. Machine Learning, 20(3):273–297,
1995.
 5. N. Eén and N. Sörensson. An extensible sat-solver. In Proceedings of the Sixth Interna-
tional Conference on Theory and Applications of Satisfiability Testing (SAT), pages 502–518.
Springer, 2003.
6. S. Haim and T. Walsh. Online estimation of sat solving runtime. In Proceedings of
the Eleventh International Conference on Theory and Applications of Satisfiability Testing
(SAT), pages 133–138. Springer, 2008.
7. F. Hutter, D. Babić, H. H. Hoos, and A. J. Hu. Boosting Verification by Automatic Tuning
of Decision Procedures. In Proceedings of Formal Methods in Computer Aided Design
(FMCAD’07), pages 27–34, Washington, DC, USA, 2007. IEEE Computer Society.
8. F. Hutter, Y. Hamadi, H. H. Hoos, and K. Leyton-Brown. Performance prediction and au-
tomated tuning of randomized and parametric algorithms. In Proceedings of the Twelfth In-
ternational Conference on Principles and Practice of Constraint Programming (CP), pages
213–228, 2006.
9. F. Hutter, H. H. Hoos, and T. Stützle. Automatic algorithm configuration based on local
search. In Proceedings of the 22nd AAAI Conference on Artificial Intelligence, pages 1152–
1157. AAAI Press, 2007.
10. M. G. Lagoudakis and M. L. Littman. Learning to select branching rules in the dpll procedure
for satisfiability. In In LICS/SAT, pages 344–359, 2001.
11. M. W. Moskewicz, C. F. Madigan, Y. Zhao, L. Zhang, and S. Malik. Chaff: Engineering
an efficient sat solver. In Proceedings of the 38th conference on Design automation (DAC),
pages 530–535, 2001.
12. E. Nudelman, A. Devkar, Y. Shoham, K. Leyton-brown, and H. Hoos. Satzilla: An algorithm
portfolio for sat. In Proceedings of the Seventh International Conference on Theory and
Applications of Satisfiability Testing (SAT), pages 13–14, 2004.
13. S. A. Seshia. Adaptive eager Boolean encoding for arithmetic reasoning in verification. Tech.
Rep. CMU-CS-05-134, School of Computer Science, Carnegie Mellon University, 2005.
14. L. Xu, F. Hutter, H. H. Hoos, and K. Leyton-Brown. SATzilla07: The design and analysis of
an algorithm portfolio for SAT. In Proceedings of the Thirteenth International Conference
on Principles and Practice of Constraint Programming (CP), pages 712–727, 2007.
15. E. Zarpas. Benchmarking sat solvers for bounded model checking. In Proceedings of the
Eighth International Conference on Theory and Applications of Satisfiability Testing (SAT),
pages 340–354, 2005.